Using reflect to characterize decision procedures

Thus far in Verified Functional Algorithms we have been using

• propositions (Prop) such as a<b (which is Notation for lt a b)
• booleans (bool) such as a<?b (which is Notation for ltb a b).

The Perm chapter defined a tactic called bdestruct that does case analysis on (x <? y) while giving you hypotheses (above the line) of the form (x<y). This tactic is built using the reflect type and the blt_reflect theorem.

The name reflect for this type is a reference to computational reflection, a technique in logic. One takes a logical formula, or proposition, or predicate, and designs a syntactic embedding of this formula as an "object value" in the logic. That is, reflect the formula back into the logic. Then one can design computations expressible inside the logic that manipulate these syntactic object values. Finally, one proves that the computations make transformations that are equivalent to derivations (or equivalences) in the logic. The first use of computational reflection was by Goedel, in 1931: his syntactic embedding encoded formulas as natural numbers, a "Goedel numbering." The second and third uses of reflection were by Church and Turing, in 1936: they encoded (respectively) lambda-expressions and Turing machines. In Coq it is easy to do reflection, because the Calculus of Inductive Constructions (CiC) has Inductive data types that can easily encode syntax trees. We could, for example, take some of our propositional operators such as and, or, and make an Inductive type that is an encoding of these, and build a computational reasoning system for boolean satisfiability. But in this chapter I will show something much simpler. When reasoning about less-than comparisons on natural numbers, we have the advantage that nat already an inductive type; it is "pre-reflected," in some sense. (The same for Z, list, bool, etc.) Now, let's examine how reflect expresses the coherence between lt and ltb. Suppose we have a value v whose type is reflect (3<7) (3<?7). What is v? Either it is

• ReflectT P (3<?7), where P is a proof of 3<7, and 3<?7 is true, or
• ReflectF Q (3<?7), where Q is a proof of ~(3<7), and 3<?7 is false.

In the case of 3,7, we are well advised to use ReflectT, because (3<?7) cannot match the false required by ReflectF.

So v cannot be ReflectF Q (3<?7) for any Q, because that would not type-check. Now, the next question: must there exist a value of type reflect (3<7) (3<?7) ? The answer is yes; that is the blt_reflect theorem. The result of Check blt_reflect, above, says that for any x,y, there does exist a value (blt_reflect x y) whose type is exactly reflect (x<y)(x<?y). So let's look at that value! That is, examine what H, and P, and Q are equal to at "Case 1" and "Case 2":

Here is another proof that uses inversion instead of destruct. The ReflectF case is eliminated automatically by inversion because 3<?7 does not match false.

The reflect inductive data type is a way of relating a decision procedure (a function from X to bool) with a predicate (a function from X to Prop). The convenience of reflect, in the verification of functional programs, is that we can do destruct (blt_reflect a b), which relates a<?b (in the program) to the a<b (in the proof). That's just how the bdestruct tactic works; you can go back to Perm.v and examine how it is implemented in the Ltac tactic-definition language.

Using sumbool to Characterize Decision Procedures

An alternate way to characterize decision procedures, widely used in Coq, is via the inductive type sumbool. Suppose Q is a proposition, that is, Q: Prop. We say Q is decidable if there is an algorithm for computing a proof of Q or ~Q. More generally, when P is a predicate (a function from some type T to Prop), we say P is decidable when forall x:T, decidable(P). We represent this concept in Coq by an inductive datatype:

Let's consider sumbool applied to two propositions:

A value of type sumbool (3<7) (3>2) is either one of:

• left applied to a proof of (3<7), or
• right applied to a proof of (3>2).

Now let's consider:

A value of type sumbool (3<7) (2>3) is either one of:

• left applied to a proof of (3<7), or
• right applied to a proof of (2>3).

But since there are no proofs of 2>3, only left values (such as v2a) exist. That's OK. sumbool is in the Coq standard library, where there is Notation for it: the expression {A}+{B} means sumbool A B.

A very common use of sumbool is on a proposition and its negation. For example,

That expression, forall a b, {a<b}+{~(a<b)}, says that for any natural numbers a and b, either a<b or a>=b. But it is more than that! Because sumbool is an Inductive type with two constructors left and right, then given the {3<7}+{~(3<7)} you can pattern-match on it and learn constructively which thing is true.

Suppose there existed a value lt_dec of type t4. That would be a decision procedure for the less-than function on natural numbers. For any nats a and b, you could calculate lt_dec a b, which would be either left ... (if a<b was provable) or right ... (if ~(a<b) was provable). Let's go ahead and implement lt_dec. We can base it on the function ltb: nat -> nat -> bool which calculates whether a is less than b, as a boolean. We already have a theorem that this function on booleans is related to the proposition a<b; that theorem is called blt_reflect.

It's not too hard to use blt_reflect to define lt_dec

Another, equivalent way to define lt_dec is to use definition-by-tactic:

Warning: these definitions of lt_dec are not as nice as the definition in the Coq standard library, because these are not fully computable. See the discussion below.

sumbool in the Coq Standard Library

The output of Print sumbool explains that the first two arguments of left and right are implicit. We use them as follows (notice that left has only one explicit argument P:

Now, let's use le_dec directly in the implementation of insertion sort, without mentioning ltb at all.

Exercise: 2 stars (insert_sorted_le_dec)

Look at the proof state now. In the first subgoal, we have above the line, Hle: a <= x. In the second subgoal, we have Hgt: ~ (a < x). These are put there automatically by the destruct (le_dec a x). Now, the rest of the proof can proceed as it did in Sort.v, but using destruct (le_dec _ _) instead of bdestruct (_ <=? _).

☐

Decidability and Computability

Before studying the rest of this chapter, it is helpful to study the ProofObjects chapter of Software Foundations volume 1 if you have not done so already. A predicate P: T->Prop is decidable if there is a computable function f: T->bool such that, forall x:T, f x = true <-> P x. The second and most famous example of an undecidable predicate is the Halting Problem (Turing, 1936): T is the type of Turing-machine descriptions, and P(x) is, Turing machine x halts. The first, and not as famous, example is due to Church, 1936 (six months earlier): test whether a lambda-expression has a normal form. In 1936-37, as a first-year PhD student before beginning his PhD thesis work, Turing proved these two problems are equivalent. Classical logic contains the axiom forall P, P \/ ~P. This is not provable in core Coq, that is, in the bare Calculus of Inductive Constructions. But its negation is not provable either. You could add this axiom to Coq and the system would still be consistent (i.e., no way to prove False). But P \/ ~P is a weaker statement than {P}+{~P} , that is, sumbool P (~P). From {P}+{~P} you can actually calculate or compute either left (x:P) or right(y: ~P). From P \/ ~P you cannot compute whether P is true. Yes, you can destruct it in a proof, but not in a calculation. For most purposes its unnecessary to add the axiom P \/ ~P to Coq, because for specific predicates there's a specific way to prove P \/ ~P as a theorem. For example, less-than on natural numbers is decidable, and the existence of blt_reflect or lt_dec (as a theorem, not as an axiom) is a demonstration of that. Furthermore, in this "book" we are interested in algorithms. An axiom P \/ ~P does not give us an algorithm to compute whether P is true. As you saw in the definition of insert above, we can use lt_dec not only as a theorem that either 3<7 or ~(3<7), we can use it as a function to compute whether 3<7. In Coq, you can't compute with axioms! Let's try it:

Now, can we use this axiom to compute with?

That doesn't work, because an if statement requires an Inductive data type with exactly two constructors; but lt_dec_axiom_1 i j has type i<j \/ ~(i<j), which is not Inductive. But let's try a different axiom:

This typechecks, because lt_dec_axiom_2 i j belongs to type sumbool (i<j) (~(i<j)) (also written {i<j} + {~(i<j)} ), which does have two constructors. Now, let's use this function:

This compute didn't compute very much! Let's try to evaluate it using unfold:

It is dangerous to add Axioms to Coq: if you add one that's inconsistent, then it leads to the ability to prove False. While that's a convenient way to get a lot of things proved, it's unsound; the proofs are useless. The Axioms above, lt_dec_axiom_1 and lt_dec_axiom_2, are safe enough: they are consistent. But they don't help in computation. Axioms are not useful here.

Opacity of Qed

This lemma prove_with_max_axiom turned out to be provable, but the proof could not go by computation. In contrast, let's use lt_dec, which was built without any axioms:

Unfortunately, even though blt_reflect was proved without any axioms, it is an opaque theorem (proved with Qed instead of with Defined), and one cannot compute with opaque theorems. Not only that, but it is proved with other opaque theorems such as iff_sym and Nat.ltb_lt. If we want to compute with an implementation of lt_dec built from blt_reflect, then we will have to rebuild blt_reflect without using Qed anywhere, only Defined. Instead, let's use the version of lt_dec from the Coq standard library, which is carefully built without any opaque (Qed) theorems.

The Coq standard library has many decidability theorems. You can examine them by doing the following Search command. The results shown here are only for the subset of the library that's currently imported (by the Import commands above); there's even more out there.

The type of list_eq_dec is worth looking at. It says that if you have a decidable equality for an element type A, then list_eq_dec calculates for you a decidable equality for type list A. Try it out:

Exercise: 2 stars (list_nat_in)

Use in_dec to build this function.

☐ In general, beyond list_eq_dec and in_dec, one can construct a whole programmable calculus of decidability, using the programs-as-proof language of Coq. But is it a good idea? Read on!

Advantages and Disadvantages of reflect Versus sumbool

I have shown two ways to program decision procedures in Coq, one using reflect and the other using {_}+{~_}, i.e., sumbool.

• With sumbool, you define two things: the operator in Prop such as lt: nat -> nat -> Prop and the decidability "theorem" in sumbool, such as lt_dec: forall i j, {lt i j}+{~ lt i j}. I say "theorem" in quotes because it's not just a theorem, it's also a (nonopaque) computable function.
• With reflect, you define three things: the operator in Prop, the operator in bool (such as ltb: nat -> nat -> bool, and the theorem that relates them (such as ltb_reflect).

Defining three things seems like more work than defining two. But it may be easier and more efficient. Programming in bool, you may have more control over how your functions are implemented, you will have fewer difficult uses of dependent types, and you will run into fewer difficulties with opaque theorems. However, among Coq programmers, sumbool seems to be more widely used, and it seems to have better support in the Coq standard library. So you may encounter it, and it is worth understanding what it does. Either of these two methods is a reasonable way of programming with proof.